OpenAI turns AI proofs over to experts
OpenAI's models produced proofs of open problems in mathematics, computer science, and physics, and outside collaborations are now shepherding the results. — openai.com
Claude 3 Opus shows emergent wisdom
Claude 3 Opus exhibits a form of emergent wisdom that makes AI alignment largely a political problem rather than a technical one. — lesswrong.com
C++26 draft fails approval ballot
The C++26 standard draft failed its approval ballot, leaving ratification of the standard in doubt. — linkedin.com
FDT hinges on undefined logical counterfactuals
Functional decision theory depends on computational logical counterfactuals, a concept nobody has defined, leaving the theory ill-specified. — lesswrong.com
New LoRAs bring editing to LTX 2.3
Newly trained low-rank adapter models bring inpainting, reference-guided editing, and experimental video editing to LTX 2.3. — huggingface.co
Ghostty reaches full Kitty graphics support
Ghostty is now the first terminal other than Kitty to implement the complete Kitty graphics protocol, with large performance and memory improvements. — ghostty.org
Control monitors need alignment checks too
Systems that monitor AI agents can fail through misalignment rather than weakness, so they need the same kind of alignment assessments as the agents themselves. — lesswrong.com
Genie 4 simulates interactive worlds with audio
Genie 4 is a real-time world model that generates interactive scenes with native audio and is now in a trusted-tester preview. — deepmind.google
AI solves open conjecture in minutes
An AI coding assistant solved an open mathematics conjecture, and a human then spent days verifying the result. — lesswrong.com
Imagination Library passes 343 million books
Dolly Parton's Imagination Library has gifted more than 343 million free books to children around the world. — imaginationlibrary.com